P64 warp-epoch cancellation
A larger P64 launch can reduce epoch-check overhead while remaining responsive enough for stale-work cancellation.
Diagnostic work, exact algebra and local capabilities are not treated as end-to-end mining advantage.
What was tested?
A larger P64 launch can reduce epoch-check overhead while remaining responsive enough for stale-work cancellation.
Why the test is meaningful
Larger kernels amortize control instructions but lengthen the interval before the device can observe a changed epoch, increasing sensitivity to copy-stream contention.
retention=throughput(P64-epoch)/throughput(P64)response=t_observed−t_signaltechnical gate: all responses≤12 ms; hard ceiling 15 msHow it was tested
Run exact P64 with warp-broadcast epoch checks, full 2^32 validation, balanced throughput comparisons and eight cancellation trials using the registered copy path.
What happened
Retention was 0.998898 with CI95 0.998552–0.999245. Responses ranged 4.77–17.61 ms: four exceeded 12 ms and three exceeded 15 ms.
Exactness and statistical controls
Exactness passed with zero discrepancies, 44 registers and zero spills. The measured jitter failed the frozen technical audit.
What the result means
P64 trades too much cancellation responsiveness for amortization. High-priority copies do not guarantee timely observation under this workload.
Limitations
- The result includes contention from the registered H2D control path.
- Tail behavior may differ on other drivers or GPUs.
- No end-to-end mining benefit was tested.
Evidence trail
Repeat full-domain P64 validation and all eight timestamped cancels; evaluate the 12 ms gate and 15 ms ceiling separately.
Canonical variants
CANONICAL-EXP-169PREREGISTERED-CAMPAIGNSEALED-AUDITSource: internally audited canonical reports. Local filesystem structure, private headers and operational identifiers are excluded from publication.